Every forecast of the AI buildout quotes the interconnection queue as though it were a pipeline. It is a waiting list, and most of it never happens. I put a calibrated probability on every request still in it, published before the outcomes exist, with a date each one must settle by.
MISO, CAISO, NYISO and ISO-NE. Every figure below reproducible from public sources.
Not a forecast of prices, and not a view on who wins. A measurement of whether physical infrastructure gets built — which is worth the same to someone long the trade as to someone short it.
The claim is narrow on purpose: when it says 20%, that should happen about 20% of the time. That is checkable, it resolves on its own, and the ledger is public so it can be checked against me. Grid operators publish current queue state and overwrite it in place, so the revision history is captured every morning — nobody can go back and buy the past.
Because the last time I measured my own work, it failed, and I published that instead of burying it. TirraMind collected 34 public data sources and 375,657 observations to detect market anomalies. Then I ran the experiment almost nobody runs and tested whether it predicted anything. Zero of fifty-one hypotheses survived Benjamini–Hochberg correction.
I reported the study's own statistical power — about 10% — up front, because overclaiming a null is the same error as overclaiming an edge. Along the way the pipeline lied to me thirteen documented times while reporting green. A ranking I was about to sell was 90% Unix timestamp.
So the signal product was withdrawn and the failure log published. The habit that killed it is the one thing worth keeping, and it is what runs the work above: check whether the number means what it is being used to mean, before anyone bets on it. The queue model publishes its own calibration failure for the same reason.